PR Docket

Inspiration

Code reviews are essential, but reviewers often focus on different concerns and may miss security, performance, or readability issues. We built PR Docket to provide every pull request with a fast, focused AI review team before it is merged.

What it does

PR Docket accepts a Git diff or GitLab merge request URL and assigns it to three specialized AI reviewers:

Security
Performance
Readability

A fourth judge agent evaluates their findings, removes duplicates, resolves conflicts, and produces one clear, actionable review.

How we built it

We built PR Docket as a stateless web application using:

React and Vite for the frontend
Node.js and Express for the API
Qwen models through an OpenAI-compatible endpoint
GitLab’s merge-request diffs API
Parallel specialist agents followed by a judge-agent workflow

Large merge requests are divided into bounded batches to stay within model context limits. The system also reports reviewed and skipped files.

Challenges we ran into

Our main challenge was reviewing large changes without losing important context. We also had to:

Produce reliable, structured model responses
Eliminate duplicate or conflicting findings
Handle malformed diffs and external API failures
Safely support public, private, and self-managed GitLab projects
Balance thorough analysis with concise results

Accomplishments that we're proud of

We are proud that PR Docket:

Coordinates four AI agents in a practical workflow
Reviews multiple engineering concerns in parallel
Converts separate opinions into one coherent verdict
Supports both pasted diffs and GitLab merge requests
Handles large or unsupported files gracefully
Requires no database or complicated infrastructure

What we learned

We learned that multi-agent systems are most effective when each agent has a narrow role and a strict output format. We also discovered that the judge agent is crucial for reducing noise and turning independent findings into a useful final review.

Most importantly, we learned that context management, validation, and failure handling are just as important as model quality.